@peerbit/shared-log 13.1.0 → 13.1.2

package/src/pid.ts

@@ -1,3 +1,5 @@
+const MIN_MEMORY_HEADROOM_BALANCE_SCALER = 0.25;
+
 export class PIDReplicationController {
 	integral!: number;
 	prevError!: number;
@@ -49,16 +51,23 @@ export class PIDReplicationController {
 		let errorMemory = 0;
 
 		if (this.maxMemoryLimit != null) {
+			// Treat the configured storage limit as a ceiling, not the exact control
+			// target. A small reserve prevents discrete entry sizes and delayed checked
+			// prunes from repeatedly settling just above the hard budget.
+			const effectiveMemoryLimit =
+				this.maxMemoryLimit > 0 ? this.maxMemoryLimit * 0.95 : 0;
 			errorMemory =
 				currentFactor > 0 && memoryUsage > 0
-					? Math.max(Math.min(1, this.maxMemoryLimit / estimatedTotalSize), 0) -
-						currentFactor
+					? Math.max(
+							Math.min(1, effectiveMemoryLimit / estimatedTotalSize),
+							0,
+						) - currentFactor
 					: 0;
 			// Math.max(Math.min((this.maxMemoryLimit - memoryUsage) / 100e5, 1), -1)// Math.min(Math.max((this.maxMemoryLimit - memoryUsage, 0) / 10e5, 0), 1);
 		}
 
 		const errorCoverageUnmodified = Math.min(1 - totalFactor, 1);
-		const errorCoverage =
+		let errorCoverage =
 			(this.maxMemoryLimit ? 1 - Math.sqrt(Math.abs(errorMemory)) : 1) *
 			errorCoverageUnmodified;
 
@@ -71,14 +80,26 @@ export class PIDReplicationController {
 		// is material. This avoids oscillations around `totalFactor ~= 1`.
 		const coverageDeficit = Math.max(0, errorCoverageUnmodified); // ~= max(0, 1 - totalFactor)
 		const negativeBalanceScale =
-			coverageDeficit <= 0
-				? 1
-				: 1 - Math.min(1, coverageDeficit / 0.1); // full clamp at 10% deficit
+			coverageDeficit <= 0 ? 1 : 1 - Math.min(1, coverageDeficit / 0.1); // full clamp at 10% deficit
 		const errorFromEvenForBalance =
 			errorFromEven >= 0 ? errorFromEven : errorFromEven * negativeBalanceScale;
 
+		const hasMemoryHeadroom =
+			this.maxMemoryLimit != null && this.maxMemoryLimit > 0 && errorMemory > 0;
+		if (hasMemoryHeadroom && errorFromEvenForBalance > 0) {
+			// Coverage surplus often means another peer has not pruned yet. Do not let
+			// that transient surplus cancel a constrained peer that is still below an
+			// even share and has storage headroom to take more work.
+			errorCoverage = Math.max(errorCoverage, 0);
+		}
+
 		const balanceErrorScaler = this.maxMemoryLimit
-			? Math.abs(errorMemory)
+			? hasMemoryHeadroom
+				? Math.max(
+						Math.abs(errorMemory),
+						MIN_MEMORY_HEADROOM_BALANCE_SCALER,
+					)
+				: Math.abs(errorMemory)
 			: 1 - Math.abs(errorCoverage);
 
 		// Balance should be symmetric (allow negative error) so a peer can *reduce*
@@ -99,7 +120,7 @@ export class PIDReplicationController {
 		// TODO make these self-optimizing
 
 		let totalError: number;
-		const errorMemoryFactor = 0.9;
+		let errorMemoryFactor = 0.9;
 		const errorBalanceFactor = 0.6;
 
 		totalError =
@@ -108,6 +129,20 @@ export class PIDReplicationController {
 
 		// Computer is getting too full?
 		if (errorMemory < 0) {
+			if (
+				this.maxMemoryLimit != null &&
+				this.maxMemoryLimit > 0 &&
+				coverageDeficit > 0
+			) {
+				// When the ring is materially under-covered, shrinking a memory-limited
+				// range can increase gap-boundary assignments and make local memory usage
+				// worse, not better. Let the coverage term dominate until the floor is
+				// restored, while preserving the hard shrink behavior for zero-capacity peers.
+				errorMemoryFactor = Math.max(
+					0.2,
+					errorMemoryFactor - 0.7 * Math.min(1, coverageDeficit / 0.25),
+				);
+			}
 			totalError =
 				errorMemory * errorMemoryFactor + totalError * (1 - errorMemoryFactor);
 		}
@@ -138,7 +173,19 @@ export class PIDReplicationController {
 
 		// Calculate the new replication factor
 		const change = pTerm + iTerm + dTerm;
-		const newFactor = currentFactor + change;
+		let newFactor = currentFactor + change;
+
+		if (this.maxCPUUsage != null && this.maxMemoryLimit == null) {
+			// CPU pressure may shed surplus replicas, but it must not create a
+			// coverage gap where the network no longer has one full copy.
+			const coverageSurplus = Math.max(0, totalFactor - 1);
+			if (newFactor < currentFactor) {
+				newFactor =
+					coverageSurplus <= 0
+						? currentFactor
+						: Math.max(newFactor, currentFactor - coverageSurplus);
+			}
+		}
 
 		// Update state for the next iteration
 		this.prevError = totalError;

package/src/ranges.ts

@@ -2568,7 +2568,15 @@ export const debounceAggregationChanges = <
 					// Keep different change types for the same segment id. In particular, range
 					// updates produce a `replaced` + `added` pair; collapsing by id would drop the
 					// "removed" portion and prevent correct rebalancing/pruning.
-					const key = `${change.type}:${change.range.idString}`;
+					//
+					// Preserve each distinct replaced range as well. Adaptive replication can
+					// shrink a segment several times before the debounced rebalance runs; if we
+					// keep only the newest `replaced` range, entries from earlier wider ranges are
+					// never revisited and can stay resident after they become prunable.
+					const key =
+						change.type === "replaced"
+							? `${change.type}:${change.range.idString}:${change.range.rangeHash}`
+							: `${change.type}:${change.range.idString}`;
 					const prev = aggregated.get(key);
 					if (prev) {
 						if (prev.range.timestamp < change.range.timestamp) {

package/src/sync/simple.ts

@@ -58,6 +58,17 @@ export class RequestMaybeSyncCoordinate extends TransportMessage {
 	}
 }
 
+@variant([0, 6])
+export class ConfirmEntriesMessage extends TransportMessage {
+	@field({ type: vec("string") })
+	hashes: string[];
+
+	constructor(props: { hashes: string[] }) {
+		super();
+		this.hashes = props.hashes;
+	}
+}
+
 const getHashesFromSymbols = async (
 	symbols: bigint[],
 	entryIndex: Index<EntryReplicated<any>, any>,
@@ -109,6 +120,10 @@ const SESSION_POLL_INTERVAL_MS = 100;
 const DEFAULT_MAX_HASHES_PER_MESSAGE = 1_024;
 const DEFAULT_MAX_COORDINATES_PER_MESSAGE = 1_024;
 const DEFAULT_MAX_CONVERGENT_TRACKED_HASHES = 4_096;
+// Retry missing entry requests when the first response was lost (for example, due to
+// pubsub stream warmup). Keep it coarse-grained so we do not hammer the network under
+// large historical backfills.
+const SIMPLE_SYNC_RETRY_AFTER_MS = 10_000;
 
 const createDeferred = <T>() => {
 	let resolve!: (value: T | PromiseLike<T>) => void;
@@ -612,7 +627,14 @@ export class SimpleSyncronizer<R extends "u32" | "u64">
 				options?.skipCheck ||
 				!(await this.checkHasCoordinateOrHash(coordinateOrHash))
 			) {
-				this.syncInFlightQueue.set(coordinateOrHash, []);
+				// Track the initial sender so we can retry if the first request is lost.
+				this.syncInFlightQueue.set(coordinateOrHash, [from]);
+				let inverted = this.syncInFlightQueueInverted.get(from.hashcode());
+				if (!inverted) {
+					inverted = new Set();
+					this.syncInFlightQueueInverted.set(from.hashcode(), inverted);
+				}
+				inverted.add(coordinateOrHash);
 				requestHashes.push(coordinateOrHash); // request immediately (first time we have seen this hash)
 			}
 		}
@@ -688,6 +710,7 @@ export class SimpleSyncronizer<R extends "u32" | "u64">
 
 			const requestHashes: SyncableKey[] = [];
 			const from: Set<string> = new Set();
+			const now = Date.now();
 			for (const [key, value] of this.syncInFlightQueue) {
 				if (this.closed) {
 					return;
@@ -696,26 +719,32 @@ export class SimpleSyncronizer<R extends "u32" | "u64">
 				const has = await this.checkHasCoordinateOrHash(key);
 
 				if (!has) {
-					// TODO test that this if statement actually does anymeaningfull
-					if (value.length > 0) {
+					if (value.length === 0) {
+						// No remaining peers to ask; drop the pending key to avoid leaking.
+						this.clearSyncProcessKey(key);
+						continue;
+					}
+
+					// Ask one peer per key per loop. If a previous request is still considered
+					// "recent", wait until the retry window elapses.
+					const candidate = value[0]!;
+					const publicKeyHash = candidate.hashcode();
+					const inflightTimestamp = this.syncInFlight
+						.get(publicKeyHash)
+						?.get(key)?.timestamp;
+					if (
+						inflightTimestamp == null ||
+						now - inflightTimestamp >= SIMPLE_SYNC_RETRY_AFTER_MS
+					) {
 						requestHashes.push(key);
-						const publicKeyHash = value.shift()!.hashcode();
 						from.add(publicKeyHash);
-						const invertedSet =
-							this.syncInFlightQueueInverted.get(publicKeyHash);
-						if (invertedSet) {
-							if (invertedSet.delete(key)) {
-								if (invertedSet.size === 0) {
-									this.syncInFlightQueueInverted.delete(publicKeyHash);
-								}
-							}
+						if (value.length > 1) {
+							// Rotate for fairness across multiple possible sources.
+							value.push(value.shift()!);
 						}
 					}
-					if (value.length === 0) {
-						this.syncInFlightQueue.delete(key); // no-one more to ask for this entry
-					}
 				} else {
-					this.syncInFlightQueue.delete(key);
+					this.clearSyncProcessKey(key);
 				}
 			}
 
@@ -761,23 +790,27 @@ export class SimpleSyncronizer<R extends "u32" | "u64">
 		return this.clearSyncProcess(hash);
 	}
 
-	private clearSyncProcess(hash: string) {
-		const inflight = this.syncInFlightQueue.get(hash);
+	private clearSyncProcessKey(key: SyncableKey) {
+		const inflight = this.syncInFlightQueue.get(key);
 		if (inflight) {
-			for (const key of inflight) {
-				const map = this.syncInFlightQueueInverted.get(key.hashcode());
+			for (const peer of inflight) {
+				const map = this.syncInFlightQueueInverted.get(peer.hashcode());
 				if (map) {
-					map.delete(hash);
+					map.delete(key);
 					if (map.size === 0) {
-						this.syncInFlightQueueInverted.delete(key.hashcode());
+						this.syncInFlightQueueInverted.delete(peer.hashcode());
 					}
 				}
 			}
 
-			this.syncInFlightQueue.delete(hash);
+			this.syncInFlightQueue.delete(key);
 		}
 	}
 
+	private clearSyncProcess(hash: string) {
+		this.clearSyncProcessKey(hash);
+	}
+
 	onPeerDisconnected(key: PublicSignKey | string): Promise<void> | void {
 		const publicKeyHash = typeof key === "string" ? key : key.hashcode();
 		return this.clearSyncProcessPublicKeyHash(publicKeyHash);